57 lines
1.4 KiB
JSON
57 lines
1.4 KiB
JSON
{
|
|
"default_model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
|
|
"models": {
|
|
"meta-llama/Meta-Llama-3.1-8B-Instruct": {
|
|
"ctx": "65536",
|
|
"trust_remote": false,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768"
|
|
},
|
|
"openai/gpt-oss-20b": {
|
|
"ctx": "32768",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "8192"
|
|
},
|
|
"RedHatAI/Qwen3-14B-FP8-dynamic": {
|
|
"ctx": "32768",
|
|
"trust_remote": true,
|
|
"valid_tp": [1],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768"
|
|
},
|
|
"cpatonn/Qwen3-Coder-30B-A3B-Instruct-GPTQ-4bit": {
|
|
"ctx": "24576",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "32768"
|
|
},
|
|
"cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
|
|
"ctx": "20480",
|
|
"trust_remote": true,
|
|
"valid_tp": [2],
|
|
"max_num_seqs": "32",
|
|
"max_tokens": "16384",
|
|
"enforce_eager": false,
|
|
"env": {"VLLM_USE_TRITON_AWQ": "1"}
|
|
},
|
|
"RedHatAI/gemma-3-27b-it-FP8-dynamic": {
|
|
"ctx": "29000",
|
|
"trust_remote": true,
|
|
"valid_tp": [2],
|
|
"max_num_seqs": "32",
|
|
"max_tokens": "29000",
|
|
"gpu_util": "0.94"
|
|
},
|
|
"RedHatAI/gemma-3-12b-it-FP8-dynamic": {
|
|
"ctx": "9900",
|
|
"trust_remote": true,
|
|
"valid_tp": [1, 2],
|
|
"max_num_seqs": "64",
|
|
"max_tokens": "9900"
|
|
}
|
|
}
|
|
} |